While vulnerabilities such as jailbreaks, hallucinations, and prompt injection are now well understood in LLMs and AI agents, the security risks introduced by MCP are only beginning to emerge.

In this session, we'll explore advanced MCP attack vectors, including:

* Command injection
* Tool poisoning through manipulated tool descriptions and metadata
* Privilege escalation via chained tool access
* Persistent context attacks
* Shadowing of tool descriptions in MCP registries

Through real-world examples and demonstrations, you'll gain a deeper understanding of these emerging threats and learn practical approaches to building more secure AI agent systems.